UNIVERSITI PUTRA MALAYSIA

BACKPROPAGATION NEURAL NETWORK FOR COLOUR RECOGNITION

ABDUL AZIZ HUSSIEN AL-NAQEEB

FK 2002 49

BACKPROPAGATION NEURAL NETWORK FOR COLOUR RECOGNITION

By

ABDUL AZIZ HUSSIEN AL-NAQEEB

Thesis Submitted to the School of Graduate Studies, Universiti Putra Malaysia, in Fulfilment of the Requirement for the Degree of Master of Science

March 2002

Dedication

To the soul of my Father

2

Abstract of thesis presented to the senate of university Putra Malaysia in fulfilment of the requirement for the degree of Master of Science

BACKPROPAGATION NEURAL NETWORK FOR RGB COLOUR RECOGNITION

By

ABDUL AZIZ HUSSIEN AL-NAQEEB

March 2002

Chairman: Abdul Rahman Ramli, Ph.D.

Faculty: Engineering

Colour Image Processing (CIP) is useful for inspection system and Automatic

Packing Lines Systems. CIP usually needs expensive and special hardware as well as

software to extract colour from image. Most of CIP software use statistical methods to

extract colours and some system use Neural Network such as Counter-Propagation and

Back-Propagation .

Some researchers had used Neural Network methods to recognize colour of

Commission Intemationale de L'Ec1airage (CIE) Models either L *u *v or L *a *b.

CIE colour components need special and expensive devices to extract their

values from an image. However, this project will use RED, GREEN, BLUE (RGB)

colour components, which can be read from an image.

3

In this research, RGB values are used to represent the colour. RGB values are

used in two forms. The first form is the actual values that are used in PPM File Format

within (0,255) and the second form is normalized RGB values within (0, I ). Back­

Propagation Neural Network is used to recognize colour in RGB values.

It is found that RGB is useful when used with Neural Network and the Normalized

RGB value is faster in the learning of neural network.

4

Abstrak tesis yang dikemukakan kepada Senat Universiti Putra Malaysia sebagai memenuhi untuk ijazah Master Sains

RANGAIAN NEURAL BACKPROPAGATION UNTUK PENGECAMAN W ARNA RGB

Oleh

ABDUL AZIZ HUSSIEN AHMED AL-NAQEEB Mac 2002

Pengerusi: Abdul Rahman Ramli, Ph.D.

Faculti: Kejuruteraan

Pemproses Imej Warna (eIP) sangat berguna untuk sistem pemeriksa dan Sistem

Pembungkusan Automatik. elP selalunya memerlukan perkakasan khusus yang mahal

serta perisian untuk menyarikan warn a daripada imej. Kebanyakan perisian elP

menggunakan kaedah statistik. Terdapat juga yang mengggunakan algoritma rangkaian

neural seperti Perambatan-Berlawanan dan Perambatan-Kebelakang.

Beberapa orang penyelidik meggunakan kaedah rangkaian neural untuk mengenalpasti

warna untuk model Commission Internationale de L'Ec1airage (CIE) samaada L *u *v

or L *a *b.

Komponen warn a CIE memerIukan peranti khusus yang mahal untuk menyarikan

n i lai-nilai daripada imej . Projek ini menggunakan pendekatan yang lain iaitu dengan

menggunakan komponen warn a merah, hijau, biru (RGB), yang boleh ditentukan

dengnn bantuan komputer peribadi biasa.

Ni lai RGB digunakan untuk menyesarkan warn a daripada imej dan seterusnya

dilaksanakan a dalam dua format, iaitu Format Fai l PPM antara (0,255) dan format n i lai

5

RGB dinonnalkan an tara (0, I), setcrusnya Rangkaian Neural Perambatan-Kebelakang

diguna untuk mengenalpasti warna yang diwakili. Oi dapati RGB sangat berguna jika

digunakan bersama Rangkaian Neural dan nilai RGB yang dinonnalkan adalah lebih

cepat dalam proses pembelajarannya.

6

ACKNOWLEDGEMENTS

I would like to begin my acknowledgement by first thanking ALLAH who

creates me. I would l ike to express my thanks to Dr. Abdul Rahman Ramli who

introduced me to Colour Image processing and Neural Network and guided me in all

stages of this work; without him this work will not be here. This was followed by

inspiration from Dr. Ramlan Mahmod to further my knowledge in neural network and

apply it in the real world, also my appreciation to Mr. Razali Yaakob for his assistance.

A special thank and appreciation for Prof. Dato Dr. Shiekh Omar Abdul Rahman

for his support during my study.

A deepest thank goes to my wife and family to support me, and a big thank to my

brother Ahmed who stands behind me in my home country and takes care of my family

during my study. My special thank goes to my dearest friend Ahmed Al-jerwy for his

appreciated support.

Lastly my deepest appreciation to all those who in one way or another have

contributed to the success of my project and eventually completion of thesis.

7

I certify that an Examination committee met on 2Sll\ March 2002 to conduct the final Examination of Abdul Aziz Hussein AI-Naqeeb on his Master of Science thesis entitled "Backpropagation Neural Network for RGB Color Recognition" in accordance with Universiti Pertanian Malaysia (Higher Degree) Act 1980 and Universiti Pertanian Malaysia (Higher Degree) Regulation 198 1 . The Committee recommends that candidate be awarded relevant degree. Members for the Examination Committee are as follows

Veeraraghavan Prakash, Ph.D. Faculty of Engineering Universiti Putra Malaysia (Chainnan)

Abdul Rahman Ramli, Ph.D. Faculty of Engineering Universiti Putra Malaysia (Member) Ramlan Mahmood, Ph.D. Faculty of Computer and Infonnation Technology Universiti Putra Malaysia (Member)

RazaH Yaakob, Faculty of Computer and Infonnation Technology Universiti Putra Malaysia (Member)

SHAMS HER MOHAMAD RAMADILI, Ph.D. Professor! Deputy Dean, School of Graduate Studies Universiti Putra Malaysia

Date: 0 8 MAY 2002

8

This thesis submitted to the Senate of Universiti Putra Malaysia has been accepted as fulfillment of the requirement for the degree of Master of Science.

9

AINI IDERIS, Ph.D. Professorl Dean School of Graduate Studies Universiti Putra Malaysia

Date:

DECLARA TION

I hereby declare that the thesis is based on my original work except for quotations and citations which have been duly acknowledged. ( also declare that it has not been previously or concurrently submitted for any other degree at UPM or other institutions.

Date: 0 8 MAY 2002

10

TABLE OF CONTENTS

Page

DEDICATION 2

ABSTRACT 3

ABSTRAK 5

ACKNOWLEDGMENTS 7

APPROVAL SHEET 8

DEC LARA TION FORM 9

LIST OF TABLES 13

LIST OF FIGURES 14

CHAPTER

I INTRODUCTION 16

General Overview 1 6

Colour Pattern Recognition 17

Artificial Neural Network 19

Statement of problem 21

Thesis Objective 21

Summary of Thesis 22

II LITERTURE REVIEW 23

COLOUR IMAGE PROCESSING REVIEW 30

Introduction 30

Colour Fundamental 31

Colour Image Coding 37

Colour Models 38

ARTIFICIAL NEURAL NETWORK REVIEW 49 Introduction 49

Definition of Artificial Neural Network (ANN) 50

Basic Concepts of ANN 51

I I

Back-Propagation 64

Summary 69

III METHODOLOGY 70

Introduction 70

The Colour Image Processing Module 73

The Neural network Module 80

Back-Propagation Algorithm 83

Summary 87

IV RESULTS AND DISCUSSION 88

Introduction 88

Image processing module 89

Back-Propagation Network Module 91

The Learning Stage 91

Difficulties During the Learning 101

The Recognition Stage 103

Comparison Study 114

Summary 117

V CONCLUSION 118

Summary 118

Achievements 118

Suggestions 119

BIBLIOGRAPHY 120

APPENDICES 124

BIODATA OF THE AUTHOR 140

12

LIST OF TABLES

Table Page

3.1 RGB components in two formats: Integer and Decimal 79

3.2 The Value of Output layer nodes representing 8 colours 82

4.1 The Learning patterns and their own desired output (the integer 95 form)

4.2 The Learning patterns and their own desired output (the decimal 96 form)

4 .3 Comparison among 13 states of Learning Rate 99

4.4 The Results of recognizing eight colours 109

4.5 Comparison between researches in term of color recognition using 116 neural network

13

LIST OF FIGURES

FIGURE Page

2.1 Perceptual representation of the colour space 32

2.2 RGB Colour Cube 40

2.3 CIE Colour Triangle Space 43

2.4 Basic Colour Space in Red- Green and Yellow-Blue Model 45

2.5 A colour gamut in CIELAB appears as follows 46

2.6 (a) The node of a neural network and (b) a biological Neuron. 52

2.7 The sigmoid function 55

2.8 The Processing Element PE 56

2.9 Examples of activation functions 56

2.10 Common models of neuron with synaptic connections a hard- 57 limiting neuron (a) hard-limiting neuron (binary perceptron) and soft-limiting neuron (continuos perceptron)

2.11 Examples of Network Topology 60

2.12 Multi-layer Neural Network 63

2.13 A biased neuron 67

3.1 The Project workflow 72

3.2 RGB values as represented in PPM file format 75

3.3 Red Colour Image IN PPM format and its RGB values 76

3.4 Figure 3.4: a: Region Of Interested, b: original image and PPM 78 file which represent ROJ.

4.1 An original image with a background 92

4.2 A preprocessed image after its background is subtracted 92

4.3 The Main Menu of the BPNN Module 92

4.4 Entering BPNN Architecture file name 93

4.5 Entering number of nodes in Input, Hidden and Output Layers 94

4.6 Entering Learning Pattern 94

4.7 Entering of desired output patterns 95

4.8 Data loaded to initialize the network 97

14

4.9 Changing the Learning rate 98

4.10 Result of learning the network for decimal value 98

4.11 Testing the network with value 102

4.12 Consulting the network with RGB Components Colour 103

4.13 Consulting the Network with RGB Component 104

4.14 A verage of error within a successful learning 104

4.15 Successful learning within 26.97 seconds 108

4.16 Result of Color Recognition for RGB Components 109

4.17 Result of Green Color Recognition for RGB Components 110

4.18 Result of Red Color Recognition for RGB Components 110

4.19 Result of Cyan Color Recognition for RGB Components III

4.20 Result of White Color Recognition for RGB Components III

15

CHAPTER I

INTRODUCTION

General Overview

Since early years for industrial revolution in nineteenth century, researchers have

been looking for the best method to automate the production lines. In the middle of

twentieth century the production lines became more automated than manual but they

still need human inspector to supervise and inspect the lines. At this point, they need the

vision and intelligence, which can be only provided by human or intelligent machine

(computer). During that time the computer was under research and development, and

was very expensive to use in inspection system. In the 70's, the inspection system was

automated completely with intelligent inspection machine (computers) and robots, but

still very expensive. Eventually the cost of all technologies equipment decrease and

building an inspection system is still costly especially for small enterprises. The other

problem in vision is the colour; the inspection system in production line should consider

the colour, to get devices to work with and furthermore analysis of colour really requires

high cost devices (Brown and Dana, 1982).

From this point, the research is involved in building a cheap inspection system

using IBM-Compatible personnel computer without any additional devices apart from

1 6

IBM-Compatible PC and a video camera. The aim is to build a software module and

integrate it with cheap hardware to provide colour inspection system.

Colour recognition usually requires special hardware and complex software. In

this research the colour recognition system is presented using normal personal computer

hardware with simple and powerful software using neural network.

ROB components are used as input for the neural network. Previous research in this

field, do not use ROB components even though it is cheap, simple to extract and

validated to represent colour, so ROB is used in this research and it is found that it is

suitable for the recognition systems. (Soriano et al.,( 2001), Razali (1999), Vlad,

(1999))

Back-Propagation Neural Network (BPNN) is used to recognize colour. Even

though there are some problems in BPNN that need to be solved, but it is a very

powerful tool to recognize and classify patterns.

Colour Pattern Recognition

Pattern Recognition is part of image processing, and the image process

techniques are used to identify object of interests among images. The object of interest is

usually extracted by using statistical methods to identify the object shape and to measure

the object size. It was only for a few years ago that the colour became considered as a

17

feature to recognize the object. Most of researchers use statistical method such as

histogram to extract the main colour of an object. The histogram method is quite slow

and complex, since the histogram algorithm needs to be implemented for each image

( frame) of the on-line steam (Brown and Dana, 1982).

A colour is very important feature of an object to be identified, so a colour can be

used as stand alone feature to recognize object especially when other vectors are fixed

(constant).

Colour Science has many standards/space/module and theorems that are adopted

by organizations such as Commission Intemationale de L'Eclairage (CIE) and its sub­

organization CIELAB, which adopted CIE XYZ Standards, L u* v* and L a* b*. These

standards are usually used to represents colour in electronic and electromagnetic devices

such as TV, Video, Printers, Colour Industry, and some IT devices (Poynton, 1997).

The RGB module which is well known and standardized is used in computer to

represents colours. In fact, all computers systems represent colour in their memories

using the RGB module. The RGB module is a module that represents colour using the

three main colours, namely Red, Green and Blue. It is a reflection for actual colours in

the life. Even the human's eye retina uses three nerves to catch the three main colours

and then transfer the electro-chemical pulse to human brain where there are special

neurons cells to recognize the colours (Brown and Dana, 1982). This study attempts to

stimulate this mechanism.

18

ROB Colour model is the common used model. ROB values can be extracted

from images using personnel computer, but other Colour models such as L u* v* values

need special and expensive devices to be extracted (Poynton, 1997).

Artificial Neural Network

Artificial Neural Network (ANN) is part of Artificial Intelligence (AI), which is

developed as a simulation for human brain neural network. ANN is a network of many

simple processors (units), each possible of having a small amount of local memory. The

units are connected by communication channels ("connections"), which usually carry

numeric (as opposed to symbolic) data, encoded by any various means. The units

operate only on their local data and on the inputs they receive via the connections. The

restriction to local operation is often relaxed during training (Freeman, 1991).

ANN has so many modules of network that were developed by many researchers

and groups since the middle of last century. Some ANNs are modules of biological

neural network and some are not, but historically, most of the inspiration for the ANNs

fields came from the desire to produce an artificial system capable of carrying out

sophisticated models, perhaps "intelligent"(Freeman, 1991).

The Back-Propagation Neural Network (BPNN) which is a multi layer network

is used. BPNN has three layers, namely the input layer, hidden layer and output layer.

Each layer may have one or more nodes/units. The numbers of nodes in each layer is

19

subject to be changed from one application to another. In others words, the number of

nodes in each layer could be determined regarding to application requirements. In this

research, the colour in term of three colour components RGB need to be recognized. For

this purpose, BPNN architecture should have three unites/nodes in the input layer. For

hidden layer, number of units/nodes should be around the number of learning patterns.

For output layer, the number of units/nodes depends on the desired output. If the BPNN

recognize one colour so the output nodes will be one only, if the output" 1" the output is

desired colour, if "0" the output is not desired colour. But if BBNN recognizes 4

colours, number of nodes in output layer should be at least 2 nodes. For 8 colours, 3

output nodes are used to represent 8 colours in binary digit.

The BPNN is a powerful tool to recognize and classify objects and it sometimes

provides good result. Even though there are problems that need to be solved such as

uncontrolled behavior of Learning Rate, never learned situation during the learning

process (never reached successful learning). These problems of learning will be

presented and discussed in Chapter Four.

20

Statement of Problem

Colour object inspection system is complex and expensive. In this research a

method can be used in inspection system using IBM-Compatible PC and Video camera

is developed without any special hardware. Back-Propagation Neural network is used as

tool to recognize colours based on Red Green Blue (RGB) color space.

Studies and researches that applied neural network to recognize color have not

use standard RGB components as feeding data for neural network, some of them use

RGB histogram, Soriano et al. (2001). Some researcher used CIE models, (Razali Bin

Yaakob, 1999, Cardei, 1999, Cai.and and Goshtasby, 1999, and Chen et aI., 1997). CIE

model can be extracted from color object by special and expensive devices. It can also

be founded by converting RGB values into CIE L a * b * model through xyz model using

complicated and long mathematical model which will effect the speed of recognition.

Thesis Objective

To develop a method to classify colour using back propagation neural network as

classification tool and RGB colour model to represent the colours.

2 1

Summary of Thesis

The following is a brief description of the contents of each chapter. The next chapter

reviews literature of previous related work in the area of colour image and neural

network. It presents colour concepts, colour models and colour image representation,

then presents neural network concepts; the Perceptron, multi-layer networks, back­

propagation and historical overview.

Chapter Three focuses on methodology, first it traces an introduction about our

research and tools used then presents a model diagram for the whole system. This

chapter also describes the method and file format, which the system uses to extract ROB

components, in neural network model, the Algorithm of BPNN is presented in details.

Chapter Four is dedicated to presenting the results of the development and

running the program. All vectors that affect the speed of processing (learning) are

discussed and their solutions are presented.

The thesis is concluded with a brief recap of the entire project in Chapter Five.

The main objectives are presented and its respective achievements are discussed. This

chapter is concluded with various suggestions for improvement and further study for

future researches.

22

CHAPTER II

LITERATURE REVIEW

Soriano et ai., (2001 ) used the supervised Back-Propagation Neural Network

(BPNN) to classify coloured fluorescent image and they used the ROB colours

histogram as inputs for BPNN. They used two techniques for the major color search:

cluster mean (CM) and Kohonen's self-organizing feature map (SOFM). Classification

with SOFM-generated histograms as inputs to the classifier NN achieved the best

recognition rate (90%) for cases of normal, scaled, defocused, photobleached, and

combined images of AMCA (7-Amino-4- Methylcoumarin-3-Acetic Acid) and FITC

(Fluorescein Isothiocynate )-stained microspheres.

Razali Bin Yaakob, (1999) used Multi-layer neural network method to

recognize the colours automatically. The data that represent the colours are scanned

using Minolta Chroma Meter which is capable in changing colour into values. It offers

five different colour systems for measuring absolute chromaticity, that is, CIE Yxy,

L*a*b*, L*C*Ho, Hunter Lab and XYZ. In this study, only L*a*b* is used. Two types

of neural network were used in the early stage of study, i.e. backpropagation (BP) and

counterpropagation network (ePN), where 100 data were used as a testing data. The

results show that CPN recognized 100% of trained data and untrained data however BP

can only recognized 49% of trained data and 48% of untrained data. When the number

23

of data is increased to 808, training process required a large size memory, learning time

consuming and low percentage of recognization. To solve this problems, two combined

CPNs model were proposed. The results are much improved compared to the previous

study, whereby the percentage for trained data and untrained data are 99%.

VI ad, (1999) explores the possibility of recovering a lost color channel from an

ROB image, based on the information present in the remaining two color channels, as

well as on a priori knowledge about the statistics of sensor responses in a given

environment. Different regression and neural network recovery methods are compared

and the results show that even simple linear techniques suffice to obtain a good

approximation of the original color channel. He represents data in two method the

standard ROB and The CIE Lab coordinate. The camera provides the standard ROB, and

then ROB is converted to xyz then CIE lab, the recovery methods applied for the both

color data. Result has shown that the best is Standard ROB with Polynomial Regression.

Cai and Ooshtasby, (1999) developed a method for detecting human faces in

color images, which is described that first separates skin regions from nonskin regions

and then locates faces within skin regions. A chroma chart is prepared via a training

process that contains the likelihoods of different colors representing the skin. Using the

chroma chart, a color image is transformed into a gray scale image in such a way that the

gray value at a pixel shows the likelihood of the pixel representing the skin. An obtained

gray scale image is then segmented to skin and nonskin regions, and model faces

representing front- and side-view faces are used in a template-matching process to detect

24